Reputation-Coupled Incentives for Federated Learning: A Budget-Matched Simulation Study under Strategic Participation
Year: 2026
Type: article
Abstract: We examine reputation-coupled reward sharing in a federated-learning simulation with heterogeneous client costs, stochastic availability, and strategic populations. RePStack combines a grid-based leader budget choice, myopic mean-field effort responses, contribution estimates, and exponentially smoothed reputation; no repeated-game equilibrium is established. The reported campaign comprises 189 trials on Fashion-MNIST and a tinyshakespeare subset, with 60 clients and 40 rounds. Across 20 dataset–cost–seed blocks, RePStack has lower final monitored accuracy than contribution-only sharing (mean difference −0.0399; reported Holm-adjusted p = 0.019), and lower effective participation than flat and reputation-only sharing (differences −0.179 and −0.173; reported adjusted p = 7.6 × 10⁻⁶). These estimates pool heterogeneous workloads. Strategist–truthful utility contrasts are positive for free-riders under contribution-only and hybrid sharing (+3.13 and +2.13 budget units), but are not matched counterfactual deviation gains. Three-seed strategic tests cannot establish immunity or equivalence. The reputation-exponent ablation reports increasing welfare from 320 to 518 units and decreasing participation from 0.068 to 0.021; accuracy is not monotone. The results support a model-specific participation trade-off. Untouched-test generalization, settlement accounting, inferential outputs, and reproducibility require the original experimental records.
Source: Systemic Analytics
DOI: 10.31181/sa44202697
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7234942362
